HU231272B1 - Method for analyze of rib shadow's and determining of contour of shadows on chest radiographic images - Google Patents

Description

SS, 14, 2019

A procedure for the analysis of chest radiographs and the definition of shadow contours on chest radiographs.

A. the subject of the invention is a method that is used to determine the contours of the chest ~ basically rib contours ~ on chest x-rays. The contour ensemble is characterized by the fact that the individual contours are located in offset positions relative to each other and have different curves to varying degrees, i.e. the individual contours cannot be produced by parallel displacement of a contour. , but the contour can also be the lower and upper contour of the shadow of a rib bone.

With the addition of modern high-frequency X-ray devices, the production of high-resolution digital X-ray images created a new opportunity for chest diagnostic screening examinations. By using computer image processing using digital images, decision support systems can be created that can help the diagnosing physician speed up the creation of the primary diagnosis and increase the screening of abnormal cases! your chance.

Different types of cancer are among the most defining diseases of our time. Within this, lung cancer is one of the most common. Early, asymptomatic detection of the disease can be life-saving. In the early stages of the disease, the patient does not yet have any external signs or noticeable pain to indicate the existence of the tumor, which is why the hair test is the only chance for early detection of the existing disease. Most lung cancers detected at an early stage have a good chance of being cured in this initial stage, so early detection and treatment of the disease is important. The A. goal is therefore to create a screening method that can be applied to a large population of people, with which early-stage diseases can be detected with great certainty. Lung tumors in the coma stage appear as faint spots on X-rays. The analysis of chest X-ray images, the search for changes in the images, the effective and reliable operation of spot detection algorithms are made difficult, the results are falsified, and in many cases the analysis is thwarted by the shadows of some anatomical parts appearing on the images. Among these, the most important are the bones, especially the clavicle, the ribs, their shadows, but other shadows, the heart, the fold of skin above the collarbone, the breast, the diaphragm can also play a similar role. they are classified as anatomical noise during the analysis of digital images, thus their position and contours need to be determined in order to filter them out, i.e. the images are free from anatomical noise

SZTHHd OOOtWi

47/2, to produce de-noised or at least noise-reduced images, for more efficient image analysis, and to make a partial diagnosis. Knowledge of the outline of the lungs is also necessary for the proper functioning of the spot search algorithms.

Two main approaches to finding and removing bone shadows have emerged over the years. One first tries to find the bones and then eliminate their shadows. The other trend directly tries to suppress the bones. In the direct approach, they try to teach the system how to get only bones from the original image. Or to create an image containing only soft tissues. In general, they try to produce the image containing only bones, and then subtract this from the original image. This teaching requires so-called "dual (two different) energy" image pairs. Bones are stored in a separate image and soft tissues in a separate image. Such image pairs can only be made with a special device, with photons of other energies inserted between the two detectors with a copper filter. able to detect (M. Leng, B. van Ginekben and AMR Sehílhman: Filter learning application to suppression of bony structures from chest radiographs, Medical image analysis, 10(6):826-840. 2006), or two successively different energies images created by x-rays are detected.

The invention provides a new method for determining rib pits and rib contours. Many methods and procedures for determining rib contours have been developed and published, and patented solutions have become known.

The history of bone detection in chest x-ray images dates back to 1973, roriwaki (X I. Tortwaki, ¥, Smmega, T. Negoro and T. Fukumura; Pattern recognition of chest x-ray images. Computer Graphics and Image Processing, 2(3 ):252-271, 1973) approximated the ribs with parabolas whose vertices fall on a common line. Depending on the horizontal position, he obtained connected components from the outputs of three different filters. He labeled these based on the inner region of the lung, so that the converging areas at the edges were also separated, and then he fitted the parabolas to the borders of these areas. The weakness of the procedure is that in some cases the regions belonging to each 'rib' flow inseparably into one another, so it is not possible to implement a robust procedure with this approach.

Weschier (H. Wesehler and X Sklansky: Finding the rib cage in chest radiographs. Pattern Recognition, 9(1 ):21-30, 1977) already tried to find the complete ribs (both the abdominal and dorsal parts) in 1976. He divided the ribs into parabolic and elliptical sections. He tried to localize these sections separately with a modified Hough transformation. A high-pass filter was used as pre-processing. The abdominal and dorsal sections are one

He paired it with 47/3 fourth degree poHnom. Using this method, 15% of the ribs found were wrong.

Souza (P.de Souza: Automatic rib detection in chest radiographs. Computer vision, graphics and image processing, 23(2):129-161, 1983) came up with a new approach in 1982. It does not apply an edge enhancement filter, so the process can be accelerated. As a first step, the lung area is segmented, so you could limit the search for ribs within the lung area. Within the lung area, vertical intensity profiles were determined and matched to each other by offset. He summed up the fitted profiles and looked for local extreme values on them. From the extreme values found, it was possible to find the other points of the ribs, which were then refined pixel by pixel. To make the output curves not too noisy, he applied a combination of two parabolas to them. With this method, the middle 5-6 ribs can be found quite quickly, but finding the other ribs is not guaranteed. One of the main disadvantages of the method is that the local extreme value search and the pixel-level refinement are rule-based and rather complicated, and the other disadvantage is that it only uses an offset when matching the vertical intensity profiles.

The big breakthrough was the 1995 article by Yue (Z. Yue, A. Goshtasby and L. V. Ackerman: Automatic detection of rib border in chest radiographs. Medical Imaging, IEEE Transactions on, 14(3):525-536, 1995). This procedure also tried to delimit the lung area for the first time by searching for the boundary line of the lung area using dynamic programming on horizontal profiles. He used the result to match the outer section of the ribs to this. He divided the rib arches, approximated the outer and inner parts with separate parabolas, and attached the inner part to an approximated inner lung border. He tried to localize these parabolas with a modified Hough transformation, which took these boundary conditions into account and used only the points with a gradient satisfying them. He filtered the resulting rib candidates with a rule-based system and then refined them with an active contour (M.Kass, A. Witkin and D. Terzopolus Snakes: Active contour models. International journal of computer vision, 1(4):321-331, 1988) procedure the result. The weak point of the method is also the rule-based system.

Ginneken's 2000 article (B. van Gínekken and B. M. ter Haar Romeny: Automatic delineation of ribs in frontal chest radiographs. In Proceedings of SPICE, volume 3979, page 825, 2000), in which he tests the posterior rib arches of the entire rib system, was a major step forward. describe with a statistical model. The lower and upper curves of the ribs are approximated by a parabola, and 9 ribs are taken into account. so with 72 parameters, you can describe the whole

47/4 rib system. It locates the lower 5-6 ribs with a relatively high degree of certainty, but there are large inaccuracies in the outline of the upper ribs,

Suzuki (K. Suzuki. H< Abe, H. MacMahon: Image-processing technique for suppressing ribs in chest radiographs by means of massive training anifical neural network. Medical Imaging, IEEE Transactions on, 25(4):252-271, 2006 ) tried to define and remove the bones with the help of neural networks. Laplace pyramid representation) trained neural networks with levels of pairs of images. The input was an image showing both bones and soft tissues, and the output was the bony image. Each neural network learned to map image elements with different frequencies. Even in the best case of the final result obtained with the method, the bones can still be easily discerned, although from the point of view of looking for spots (recognition of lung changes), this can already help the diagnosis.

Also in 2006, Loog and Ginneken (M. Loog, B. van Ginekké»; Segmentation of the posterior ribs in chest radiographs using .iterated contextual pixel classification. Medical Imaging, IEEE Transactions on, 25(5):602-611, 2006 ) came up with a completely new approach, they proposed a method similar to direct rib elimination. For some pixels of the image containing bone, the intensity values of the same pixel of the bone-free image were not calculated, but only to say whether the given pixel belongs to bone or not. Their results are moderately successful, almost similar results can be achieved with linear filters and thresholding,

A few months later, Loog and Ginekken (M, Loug. B« van Ginekken and AMR Schilham: Filter learning application to suppression of bony structures from chest radiographs. Medical image analysis, 10(6):826-.840, 2006) also came up with an idea came up, which does not want to map the entire image into a lumpy image, but maps independent pixels with the help of a filter. Vectors were generated from the neighborhood of each pixel using general operators. These vectors were compressed, and then the average of the output corresponding to the training vector closest to the input vector was assigned to the output. The method eliminated the bones better than the previous ones, but the lung tissue was also greatly damaged.

Park (M. Park, X S> Jm, X. S, Wilson: Detection of abnormal texture in chest x-rays with reduction of ribs. In proceeding, of the Pan-Sydney area workshop on Visual information processing, pages 71-74. Australian Computer Society, Inc. 2004) first segmented the ribs and then extracted their shadows. He tried to solve shadow extraction simply. On vertical intensity profiles, the difference between inside and outside the rib was subtracted from inside the rib. With the method, the lung tissue is not scarred at all, the shadow of the ribs is greatly reduced, but the edges of the ribs remain.

47/5

Vogelsang (F. Vogelsang, F. Weiler, X Dahmen, M. Kilbmger, B. Wein and R. Gimther Detection and compensation of rib structures in chest radiographs for diagnosis assistance. In Proceedings of International Symposium on Medical Imaging, volume 3338, page L CCiteseer, 1998) already used a rib elimination method that did not compensate uniformly along the vertical lines, but tried to determine the intensity of the rib at a given point by taking into account the side columns as well.

Simkó (G, Simító, G. Orbáo, F. Máday and G, Horváth: Elimination of clavicle shadows to help automatic lung nodule detection on chest radiographs. In 4th European Conference of the International Federation for Medical and Biological Engineering, pages 488-491. Springer, 2009) uses a more sophisticated method. The vertical intensity differentiated profits, washed them both vertically and parallel to the arc of the bones, and then this was subtracted by the invisible boh. With this method, it is possible to reach a level where the damage to the lung tissue is not yet noticeable to the naked eye, while the location of the bones can no longer be identified.

It is a general opinion that the entire rib system was first detected by Plourde (F. Plourde, F. Cheriet and X Dansereau: Semi-automatic detection of scoliotic rib borders from postero-anterior chest radiographs. Biomedical Engineering, IEEE Transactions on, (99):1-1 , 2011) was found by a semi-automatic method, which requires 4 points per rib from the user for segmentation. Plourde tries it with the following procedure starting from the starting points. to follow the line of the ribs and, taking advantage of the edge parallelism of the lower and upper ribs, selects the possible route from the edge-following graph. This solution gives hope in terms of the complete solvability of the problem, but its weakness is that it needs the points specified by the user, the accuracy of which significantly affects the accuracy of the method.

US 8,094,904 B2 (Method and system for bone suppression based on single x-ray image. Gregory G. Slabaugh, London (GB), Tong Fang, Morganville NJ (US), Kooksang Moon, North Brunswick NJ (US), Yunqlang Chen Plainsboro NJ (US) 12.09.2008) patented procedure starts from a single internal X-ray recording and realizes the suppression of bone structures, i.e. the production of an image without bone structure, containing only soft parts. As a first step, 48 different filters (Leug-Malik, Gauss filters) basically filter out components of different frequencies on the pixel level. After that, the image containing only soft parts is formed by starting from the original image and determining it at the pixel level using learning regression functions (GPR, KNNR). intensity values for a given pixel for an image containing only soft parts, .The student is regressive

Functions 47/6 use the features filtered at the pixel level from the input x-ray image at the beginning of the process.

The subject of the patent WO2011077334 Al (Bone suppression in x-ray radiograms. Berg Jens Von, Ulrich Neitzel 2011,06,30) is a system that can generally be used to remove an object (rib bone) bounded by a contour from a source image (x-ray). The first element of the method is a gradient unit, which calculates a gradient field from the source image. This is followed by an equalizing / smoothing / blurring filter, which shapes and refines the gradient field. The following integration unit a. creates the image of the object to be removed (rib bone) by integration from a modified gradient field. Finally, the rib bone image is extracted from the original image. The filtering is a two-dimensional convolutional kernel, which is a combination of a one-dimensional convolutional kernel parallel to the rib contour and a one-dimensional convolutional kernel perpendicular to the rib contour. The first kernel (parallel to the rib contour) filters hair along the rib contours on both sides of the contour, while the second kernel filters perpendicular to the contours in the areas separated by the contour, regardless of the position and curvature of the contour.

The invention WO2014954018 A2 (Bone suppression in x-ray imaging, Andrae Goossen, Raoul Florent, Claire Levrier, Berg teas Von 10.04.2014) is basically implemented for medical image viewing applications, including mainly when viewing cardiology, angiography images or screening images and partial bone suppression. The system consists of three parts. It includes an image data input unit, a processing unit and a display unit. A. the image data input unit is "set up" so that it only receives information from a region of the image (cardiology, angiography) designated by the user. performs suppression (partial suppression) in the designated range. The display unit is "set up" to display the partially occluded image. The patent does not detail what the setting of each unit entails, it does not provide information about the specific operation, and it is claimed that it can only achieve partial bone suppression.

The patent 20140079309 (Rib suppression in radiographic images, Zhimin Huo (Pittsford, NY, US), Jing Zhang (Shanghai, CN. 2014,03.20)) implements rib detection and suppression in chest X-ray images. The method first basically looks for rib contours (one or more ribs) on the X-ray image with an edge highlighting / edge search process, which it labels. After that, the found rib arch selects the best fitting one from the rib system models in the database. Using the selected model, it refines the starting rib contours and searches for and fits additional rib contours. Meanwhile, he modifies the

47/7 selected. parameters of the rib system model, taking into account the individuality of the patient. After delimiting the interactive rib, it forms the suppressed image of the rib. The method requires a large number of rib system models in advance.

The "image component separation method" according to patent US 8,155,412 (Device, method and computer readable recording medium containing program for separating Image components. Kitamura; Yoshiro AshigarakamLgun, JP) generally realizes the separation of more than one predetermined image component by using more than one X-ray image. . When defining and suppressing the thoracic rib system, more than one image means two x-rays with different energies, and more than one component means the rib system and soft parts (lungs). To extract the required image component from images with different energies, it performs a weighted summation on the pixels of the image using predetermined weighting factors. To apply the method, two recordings with different energies are required.

Patent 1187,545,965 (Image modification and detection using massive training artificial neural networks (MTANN.) Suzuki; Kenji (Clarendon Hills, IL), Dol; Kunio (Willowbrook. IL) 11/10/2003) offers a method, process and computer program in general, to modify the anatomical structures appearing on a doctor's diagnostic image (x-ray) (suppression of the skeletal system). The procedure requires a so-called first medical image that includes the entire anatomical structure. After that, the first medical image is processed multiple times. The resulting images are created by learning artificial neural networks (MTANN). Neural networks are trained to display the anatomical structures of a given resolution (given frequency range) of the first image. By combining the images produced in this way, the so-called second medical image. If the task is to suppress the bone structure, then the images with different resolutions (with frequency content) are the frequencies belonging to the bones, the resulting so-called the second image is the bony image, which, when subtracted, is the so-called the printed image is created from the first medical image.

US 8,913,817 (Rib suppression in radiographic images. Huo; Zhimin Pittsford NY US, Xu; Fan Pudong Ν Ά CN, Chen; Shoupu Rochester NY US, Chen; Qinran Pudong N/A CN 2014-12-06) patent procedure recommendation chest x-ray for rib compression. The method recognizes and labels one or more ribs in the examined area and determines the rib edges on the recognized ribs. It then generates transverse rib profiles along the recognized rib. Finally, it modifies the original X-ray image using the defined (at least one) rib edges and rib profiles. The method first decides at the pixel level whether the given pixel belongs to a rib bone or not. Then he thought it belonged to the rib

On 47/8 pixels, it defines ribs using some previously known rib structure (rib wedge rib thickness, rib position). The method requires a starting database.

The method according to patent 201.40140603 (Clavicle suppression in radiographic images. Huo; Zhimin 22.05.2014,) realizes the search for and suppression of the clavicle on an X-ray image. It first defines the lung region, then at least a part of the clavicle that is outside the lung region. For the search, it uses an edge highlighting (differentiating) filter, or a learning system, or a gradient method, or a Laplace filter. After that, it determines the edges (contours) of the former part of the clavicle, and by tearing out its fibers, it also determines the edges of the clavicle in the region inside the lungs using an edge-finding procedure. After that, generate the mask of the clavicle bounded by the edges.

All rib-finding procedures use a number of similar steps. Such steps include segmenting the lung area, determining the initial contour of the ribs, and then gradually refining the rib contours. The shortcoming of the Known procedures is that either a mathematical model (ph parabola fitting) is used that is not able to handle the unique variety / difference of the rib contours, or if it is handled, it is handled incorrectly. While each rib has unique curvature contours, a joint model of the entire rib system must also be created and used, which the above procedures do not apply at all or not properly. Another shortcoming is that in the formulated optimization task, an inefficient, rule-based solution is often used, so the solutions so far are not robust and/or accurate enough, and they only find a part/fragment of the rib contours on the lung area. Therefore, the goal can be formulated to develop a more efficient, faster, more accurate and more effective procedure than before.

The essence of the solution according to the invention is the following findings:

The essence of the solution according to the invention is primarily the realization that by using a two-dimensional directional field with a density matching the rib system, including the rib thicknesses and inter-rib distances, which faithfully reproduces the tangents of the rib contours, or more precisely the slope of the tangents, the assumed rib system ribs determined by the center line from the center line points, per started step! by integration, the rib contour can be determined - in general, without knowing the mathematical function describing the rib contour.

The essence of the solution according to the invention is, secondly, the realization that the method that can be advantageously used to find a rib arc belonging to a pixel p is the parameters of matching circular arcs intersected in the middle by a short vertical section passing through pixel ap, such as the radius of the matching circular arc R, the curvature of the matching circular arc κ (R ^{::: :} ί/κ), the fitting circular arc

47/9 tilt angle, s vertical displacement of a matching circular arc, complete traversal between fixed extreme values with a given resolution, until the difference of the intensity values above and below the circular arc summed up on both sides of the circular arc, in a band of width d^, reaches its maximum value in the case of searching for the lower arc of the rib. After finding the local maximum value of the cost function that also uses the above intensity values, the first approximation value of the direction field in the analyzed pixel p is given by the direction of the tangent of the corresponding circular arc in radians. When searching for the upper arch of the ribs, the minimum of the difference between the intensity values above and below the matching circular arch must be sought, since there is no rib above the upper rib arch, and the intensity values there are small. The specific ranges of the parameters of the fitting circular arc, the bandwidth d^, were determined from the statistics made by analyzing a large number of real radiographs, which are given in the description of Figure 2.

The essence of the solution according to the invention is, thirdly, the recognition that, due to the feasibility of the above search, as well as the compromise between execution speed and quality, i.e., a fast algorithm, but still a manageable costal arc approximation with errors for further processing, the curvature of the fitting circular arc κ, the the set of values of the tilt angle of the fitting circular arc and the vertical displacement of the fitting circular arc s must be discretized, i.e. it must be specified how many n different values the above parameters can take between the extreme values evenly distributed on the linear scale. The set of values of the parameters and the step interval were determined from the statistics made by analyzing a large number of real radiographs, which are given in the description of Figure 2.

The essence of the solution according to the invention is primarily the realization that the method detailed so far also ensures effective interference filters. The reason for this is partly the intensity summation as integration, and partly the value limitation of the curvature of the fitting circular arc κ and the angle of inclination of that fitting circular arc.

The essence of the solution according to the invention is, in the fifth place, the recognition that the use of the average direction field when determining the broad circular direction field reduces the effect of the anatomical ZAiok. Namely, the upper arcs are filtered with the smoothed, averaged direction field generated from the lower arcs, while the lower arcs are filtered with the smoothed, averaged direction field generated from the upper arcs. For example, the averaged direction field for the lower arcs is simply produced by averaging the direction field values belonging to the lower arcs of a given lake image. . The above method is based on the fact that, in the case of a given rib, the shape and outline of the lower and upper rib arches are essentially the same, and their relative parallelism can be assumed,

Sixthly, the essence of the solution according to the invention is the realization that it is advisable to determine the rib contours in two steps. In the first step, lower and upper are possible

47/10 rib lines are determined starting from the midline points of the ribs ac _i;5 i^ by two-way integration of the direction field, then the final VBI rib lines are selected as the possible rib arcs (rib lines).

Seventhly, the essence of the solution according to the invention is the realization that the DPABPME dynamic programming-based rib arch position determination procedure, which also uses the one-dimensional rib signal and the statistical model for the likely location and thickness of the ribs, can be advantageously used to select the position of the final rib lines.

Eighthly, the essence of the solution according to the invention is the recognition that the ribs defined by the DPABPME dynamic programming based costal arch position determination method can be further refined with the DPACSPE dynamic programming based bone outline refinement method.

The essence of the solution according to the invention is primarily the realization that the accuracy of the DPACSPE dynamic programming-based bone contour refinement procedure can be improved if the procedure is applied simultaneously to the lower and upper ribs in the DPA2CSPE dynamic programming-based double bone contour refinement procedure, using the fact that in the case of a given rib the relative parallelism of the lower and upper rib arches can be assumed.

The main steps of the edge search method according to the invention are: searching for convex circular arc-shaped edges in the images, fitting a direction field to the found edges, using the direction field to generate a set of curves, which curves are possible edge edges, selecting among the curves those that fit the original image well and also according to a probability model possible carrier arcs, refinement of selected curves.

Claim 1 of the patent includes the entire process of the rib search, starting from the two-dimensional X-ray image, reaching the definition of the rib system, repeating the steps of the rib search in strict order. It describes in detail the process of the KKAÉK convex circular edge search unit according to the invention, the EDÖE one-dimensional rib mark generation unit, the DPABPME dynamic programming based rib arch position determination method, the DPACSPE dynamic programming based bone contour refinement method and the DPA2CSPE dynamic programming based double bone contour refinement method, which represent the novelties of the patent.

Claim 2 defines an expedient limitation to be applied during the process of the KKAÉK convex circular arc-shaped edge-finding unit.

Claim 3 defines filtering to be used during the process of the KKAEK convex circular arc edge search unit.

47/11

Claim 4 defines a specific implementation of the ESZÚ image prefilter unit.

For a detailed description of the method according to the invention, as well as the concepts necessary for its interpretation, we use drawings, where it is

Figure 1 illustrates the circular arc used to search for a rib arc belonging to a pixel p, as well as the parameters used for the search, the

Figure 2 shows the pixels located at a given distance on both sides of the circular arc belonging to a pixel (for n~5 cases) for the summation of the intensity values, the

Figure 3 shows the mesh used to search for the rib cage on the left lung side, a

Figure 4« shows the formation of the approximating ribs v based on the direction field, that is

Figure 5 shows the possible fitting of two different approximate rib arches to a real rib, a

6a. and the first two steps of modifying the rib signal h[n] in figures ó.b, while a

7a. and 7b. Figures show the third and fourth steps of modifying the rib signal h(n), a

8a. and 8b. figures show the quantities required for the statistical characterization of the rib system and their derivation, a

9a. and 9b. figures show rib thickness values and inter-rib distance values obtained from real samples and the polynomials superimposed on them, the

10a. and 10b. figures show the ratios of consecutive rib thicknesses obtained from real samples, as well as the ratios of consecutive inter-rib distances and the polynomials superimposed on them, the

Figure 11 illustrates the steps of the simplified rib search, a

Figure 12 shows the steps of the detailed rib search, the DPABPME dynamic programming based rib arch position determination procedure, the

Figure 13 shows the process of transitioning from the upper rib arch to the lower rib arch of the same rib during a detailed rib search, during the DPABPME dynamic programming based rib arch position determination procedure, the

14, figure is. lower rib cage! shows the process of moving to the upper rib arch of the next rib during a detailed rib search, during the DPABPME dynamic programming based rib arch position determination procedure, the

Figure 15 illustrates the DPACSPE dynamic programming-based bone outline refinement procedure, the

Figure 16 illustrates the DPACSPE dynamic programming-based bone contour refinement procedure and the interpretation of the markings for three starting curve points, extracted from Figure 15.

47/12

17a, 17b and 17c. The figures show three possible implementations of the normal direction filter applied during the DPACSPE dynamic programming-based bone outline refinement procedure, the

Figure 18 illustrates the course of the DPA2CSPE dynamic programming-based double bone contour refinement procedure, the

Figure 19 shows the structure of the three-dimensional block used to store the costs during the simultaneous fitting of both edges of the rib, the

Fig. 20 illustrates the structure of a slice of the former three-dimensional array, finally a

Figure 21 shows the detailed, block-diagram-like structure of the method according to the invention, as well as its operations.

In accordance with claim 1, the method according to the invention is implemented as follows. Figure 21 shows the order and relationship of the individual steps. During the preliminary processing, the KTRF double-sided lung X-ray image passes through the EFE pre-processing unit to the KKOB image coordinate system adjustment unit, and then to the KOEO double-sided X-ray image conversion unit. The frame of the already one-sided, pre-processed image with a coordinate system is created by the KKEB image frame setting unit, then the ESZE pre-segmentation unit determines the lung contour (Figure 3) and creates the J1 pre-segmented image. The pre-segmented image J1 obtained in this way is shaped by the ESZÜ image prefilter unit, which is part of the KE evaluation unit, using the J2 ESZÜ input parameters from the PBE parameter setting unit and generates the J3 prefiltered image.

The possible implementation of the ESZÜ image prefilter unit according to claim 4 is a conical kernel, the structure of which; (120-r)/(r+5) + 10 +1.05r +l000gauss _tt ^ (r) if |r | < 15, otherwise 0, kernel J2 input parameters are kernel radius r, μ (vertical rib) position and standard deviation σ.

After that, we look for convex circular arc-shaped edges in the pre-processed, one-sided, pre-segmented, filtered image with a coordinate system, and then apply a direction field to the found edges. Based on the direction field, we generate an army of curves (arcs), from which we then select the approximate arcs that fit the image well and, in addition, their joint location is highly probable, i.e. possible rib arcs, according to the probability model to be described later. Finally, the selected arcs (curves) are refined. Figure 1 illustrates the search for the rib arch corresponding to pixel ap. For the rib arch search, we use a circular arc in the KKAEK convex circular arc edge search unit. The input of the unit is the pre-filtered image J3, the .14 parameter ranges and applicable resolutions required for the search come from the PBE parameter setting unit (Figure 21), namely the curvature of the fitting circular arc κ, the radius of the fitting circular arc R (R - l/κ), U is the length of the matching circular arc, d^ is the width of the matching circle, the inclination angle of the matching circular arc, s is the vertical offset of the matching circular arc, and n is the number of pixels that are used to sum up the intensity values in the di/2 width band when searching for an extreme value. ap is at a vertical offset distance of the matching circular arc s. AG At the middle point of the matching circular arc, a tangent passing through the matching circular arc makes an angle of inclination with a line perpendicular to the vertical offset of the matching circular arc s passing through the middle point of the matching circular arc. In Figure 1, β is the angle traversed by the fitting circular arc. The individual parameters of the circular arc are completely traversed between the fixed extreme values of the parameter space with a given resolution. (The parameters used, their extreme values and the number of values that can be taken with an even distribution between the extreme values are determined by equations (1)...(9). Then, according to Figure 2, we create a band of width dk symmetrically along the circular arc. For each possible parameter variant, separately, we sum up the intensity values at the bottom and top of n pixels at the edge of the upper and lower band of width dy2 and form their difference, i.e. the . from the intensity values summed at the edge of the upper d^/2 width band, we subtract the intensity values summed at the edge of the lower <M2 width band and then find the extreme value of the differences, or the maximum in the case of rib bottom arc search. After finding the local maximum value of the cost function in the analyzed pixel p, the value of the energy field is given by the direction of the tangent of the circular arc giving the maximum intensity difference in Indian. In the case of point ap in Figure 2, it refers to n«=5 pixels per side, i.e. pn pts upper pixels and p _a t..p ₈ s lower pixels. The above description applies to the search for the lower rib cage. When searching for the upper arch of the ribs, the minimum of the difference between the intensity values above and below the arch must be sought, since there are nines rib content above the upper arch of the rib. The feasibility of the search, the compromise between speed and quality, i.e., a fast enough algorithm, to ensure a still manageable costal arch approximation error for the vertical displacement of the matching circular arc s, the pitch angle of the matching circular arc and k. the set of values of the curvature of a fitting circular arc must be discretized and limited. After testing the program on a large number of real radiographs, the best results, i.e. the optimal compromise between speed and quality, are ensured by the following empirically determined values and ranges:

47/14 which, as part of the J4 parameter ranges and applicable resolutions, are sent to the KKAÉK convex circular arc-shaped edge search unit from the PBE parameter setting unit (Figure 21). The numbers refer to a unit width image. H denotes the set of values that can be taken by the current parameter (indicated in the index). Based on a large number of program running experiences, the following values provide a good setting for the remaining free parameters:

n - 20(7) l _k = 0.075(8) d _k = 0.0015(9)

The KKAEK convex circular edge search unit saved the misaligned direction field J5 created by the above procedure (Figure 21). According to claim 2, the applied method also provides effective interference filtering. The basic reason for this is intensity summation as integration, as well as the above limitation of the inclination angle of the fitting circular arc and the curvature of the fitting circular arc κ.

Starting from the unaligned directional field J5 as input data, the II direction field matching unit applies the procedure of the KKAEK convex circular edge search unit to the points of the .16 hexagonal grid (Figure 21) and creates the matched directional field as Φφ) directional field function (21) in the pixels p . figure). Figure 3 shows the evenly spaced grid used to define the direction field, i.e. the points of the hexagonal grid J6, bounded by the lung contour ci, which is assumed to be known. By mirroring the right side of the chest to the left side, two images of the left half of the chest are obtained for one patient. The rib detection procedure runs independently on both halves of the chest image. Due to the mirroring of the right side of the chest to the left side, the same rib-finding procedure can be used for both sides of the lung, so the procedure will be described for the left side of the lung. The shape of the ribs is described with a guideline valid for the left half of the chest, henceforth using the name "chest". Since the ribs below each other are only slightly different and the change in their shape is continuous, it is possible to describe a chest image with a two-dimensional function. The direction field function Φ(ρ) is a function that assigns to each pixel p a tilt angle with the set of values: +π/2. The direction field is not created exactly in the selected points according to Figure 3, because during the extreme value search, the direction field points can change to a limited extent in the vertical direction by the vertical shift of the matching circular arc s. The output of the unit is the Φφ) directional field function (Figure 21), which contains the adjusted directional field values. Using the above method, the procedure determines the values of the guide field for each new image, i.e. for each patient's X-ray, according to Figure 3

47/15 starting from fixed grid points. An average direction field can be created by averaging a large number of predetermined direction fields. From the average direction field, it is possible to determine the limits and ranges of the parameters of the circular rib arch search, as well as, in the case of some parameters, the specific values. We determined the ranges and values according to equations (1)„ .(9) based on a large number of samples. In accordance with claim 3, filtering is used to eliminate false determinations, namely the upper arcs are filtered with the smoothed direction field generated from the lower circular arcs, while the lower arcs are filtered with the smoothed direction field belts generated from the upper circular arcs.

After creating the directional field, the rib arcs are determined in two steps in the EDBE one-dimensional rib signal generating unit, whose input is the Φ(ρ) field field function. In the first step, the intersection point of the lower and upper rib arches with the center line of the ac _m y ribs is determined, then the rib lines can be calculated starting from the intersection points by integrating the direction field.

Knowing the intersection point of one point of the rib arch with the center line of the ribs ac _R) ^, the entire rib arch is determined by the "integration" illustrated in Figure 4 in the EDBE one-dimensional rib signal generating unit. Figure 4 illustrates the journey from the i-th point qi of the g, i-th curve to the i+l-th point of qn and to the il-th point, with a step distance u. The q, ith point is the center line of the ribs and ag, intersection of the ith curve. The angle of inclination of that matching circular arc is q. point is the value of the directional field function Φ($) at the point q·. The step is similar in the other direction:

qu í ^::: qi * u(coso, sina) qm q. - u(«.w, sine) (10) (l1)

Among the curves of the curve group g determined by the above method, g$ is the uppermost curve, the curve of the rib closest to the head, and gN is the lowest, Nth curve. The set of generated points of the ith curve G· (not shown in the figure) contains the points of the gi curve generated according to equations (10) and (11). Intensities are assigned to these curves. The h; Intensity assigned to i-th curve:

where lk^(p) is the image intensity function, which gives the intensity of the pixel at any point p. The h; Based on the intensity values assigned to the i-th curve, we define the rib signal h[n]:

h(n] ~ h _n * ? if 0 < n < N ^::: 0 otherwise

After preprocessing, the procedure for selecting approximate rib trees uses these discrete one-dimensional h(n) rib marks. Preprocessing steps: hln] rib mark -* smoothing, hs; _m í _E [n]

47/16 smoothed ridge signal differentiation, h^fn] differentiated ridge signal normalization, Η _!Κ> π4η] normalized ridge signal —> downsampling, hpr«p[n] of undersampled ridges, which is also the output of the EDBE one-dimensional ridge signal generation unit (Fig. 21 figure). Smoothing is required before differentiation, since g. curves do not fit the real rib arches exactly enough. Therefore, some sections of a real rib arch and ag: curves can coincide in different places, generating several extreme values. This is illustrated in Figure 5. Figure 5 shows two possible approximate curves ag _a and gb, a gray field illustrating a possible location of the real rib, and the two extreme values shown by the differentiated rib signal Β^η]. Differentiation is necessary because we are looking for those values where the contour along the given g curve is the strongest. These locations will be the extreme values of the differentiated signal. Normalization is necessary so that the quality and characteristics of individual images are not affected. the procedure for selecting approximate rib arches. Undersampling speeds up the rib selection process. The effect of individual steps of the preliminary processing on the rib signal is shown in Figures 6a, 6b, and 7a, 7b. are illustrated in figures. The 6a. ah[n] rib symbol in figure 6b. in Fig. ah _S o4n| smoothed rib mark, the 7a. h&dn] is the differentiated ridge signal, and figure 7b shows the normalized ridge signal ah _nofm [n] as a continuous curve. In the figures above, the rectangular signal sequence is the segmented signal of the given processing step.

The h _p « obtained after the modifications. _f [n] down-sampled rib signal contains all the information necessary to determine the positions of the approximate ribs. All you have to do is select the n indices corresponding to the largest local extreme values, and the corresponding gi curves will be the approximate ribs. In many cases, the curves belonging to the largest local extremal edges found in the under-sampled edge signal h _w [n] provide a significant part of the real edge edges, but in general this method cannot be used by itself, mainly because there are more extreme values than the number of edge edges we are looking for, as this is shown in Fig. 5 and Fig. 7b. can also be seen in fig. Therefore, we also use statistical characteristics to select the final VBI rib arches, namely the statistical distributions of the general and relative locations of the rib arches calculated from the data of a large number of rib systems.

The quantities used in the rib system statistics are shown in Figs. 8a, 8b. is described based on figures. The 8a. and 8b. in the figures, rib ab, the rib above rib bj ab, and the rib below rib b^ can be seen in grey. The vertical coordinate is generally denoted by μ vertical rib position, namely gb-tfenöb is the position of the upper arch of the rib below the rib, μι^Β is the position of the lower arch of rib, Uh.feisö is the position of the upper arch of rib b, μ^^ b is the position of the lower arch of the rib above the rib, μ&.b is the position of the upper arch of the rib above the rib. The 8a. in the figure, Vb ^8bs is the thickness of rib b, Vh-.f ^ws b rib

Rib thickness above 47/17, 8b. in figure at^ b rib and U; The distance between the rib below the rib b and the distance between the rib ím ⁰ ** b and the rib above the b j rib can be seen. The 8a. based on the figure, the thickness of rib ab and the rib above rib bj b:

Vb ' ~ Mb,alsö “ PbJWíö(14a)

- Pw.ehó hb-I.febő (14b) to 8b. based on the figure, rib b and K? Rib below rib b and the distance between rib b and rib above rib b in b:

^::: PhH.feísb(15a) th-í ' ~ ^b,feiss Hb-i ,ais6(15b)

The ratio of the thickness of rib W ^ei av^ ^s b and the thickness of the rib above rib b.· b:

_x , fd .... iáraí h*- the distance between the rib tb ⁸⁰ * ab and the rib under the rib ab*> b, and ah _s a ⁸ ' ^? ” the ratio of the distance between rib b and the rib above rib b.· b:

tb ~tb ' feav Λ)

Using the above quantities, we define conditional density functions. .The random variable indicating the height of the upper arch of the rib b gives the value of the upper arch of the rib μι,,^, and the random variable V _a b _S rib thickness gives the thickness of the same rib. Based on these, it is possible to write down the probability that a given rib, whose upper curve is in position μ, has rib thickness va, i.e. the conditional density function for rib thicknesses;

f V _S bs (v | ™ μ) (18)

Mh, the random variable indicating the height of the lower arch of rib b with head _S gives the μ position of the lower arch of rib b, that is, the exact height of the lower rib arch μ^^, and the random variable T ₃ h _S distance between ribs in the same μ position specifies the distance between the ribs. Based on these, it is possible to write down the probability that a given rib, whose lower curve is in position μ, is at an inter-rib distance t from the neighboring rib, i.e. the conditional density function for inter-rib distances:

f T _sfe; (t | M _w> μ)(19) conditional density function for the ratio of V _re í successive rib thicknesses:

f Vrsl (v | Mbjys;r ^::: μ)(20) and the conditional density function for the ratio of successive rib distances T _fe };

Ijei (t | ~ μ)t.2l)

47/18

Conditional urgency functions can be determined using data extracted from sample images. We use real sample images on which the ribs are drawn. Based on this, the rib thickness, V** inter-rib distance, v^* the ratio of consecutive rib thicknesses, and tt^ the ratio of consecutive inter-rib distances, if rib ab has neighbors, can be calculated for each rib b, by collecting the Vh ^abs from each image rib thickness, tj/”* distance between ribs is the ratio of consecutive rib thicknesses, and tb ^rS! the ratio of successive intercostal distances gives quantities a. with their corresponding μ positions, we fit _a third-degree ^polynomial to each of the resulting sets. in the figure hd ^bs - ^{in the} figure av^' - PWebö ós 10b. figure shows the set of value pairs tb ^feí - pb^ and the curve of the fitted polynomial. The fitted polynomials: the 9a. in Fig. ν^Χμ) the polynomial fitted to the rib thicknesses, Fig. 9b. in Fig. at^Xp) polynomial fitted to rib distances, Fig. 10a. in Fig. ν^ίμ.) is the polynomial fitted to the ratio of successive rib thicknesses and in Fig. 10b<t' ^e5 (p) is the polynomial fitted to the ratio of successive rib distances. The polynomial fitting approximates the following conditional expected values:

v ^abs (p) ~ J fv _shs (v | M _Ws!s$ - μ) vdv (22) t^Cp) J f Tabs (t | MxaisS p) vdv(23)

V^(p) ~ jf (v | Mkrusö ~ μ) vdv(24) t ^! %) « J f (t IM _Mis6 - μ) vdv(25)

In order to be able to approximate the conditional density functions, we also need the variance of the above quantities. The standard deviation is less dependent on the μ position, so it is determined independently of it. The result for the σ/ ⁰ * rib thickness standard deviation:

47/19

In fact, we need this value, because it was approximated directly from the sample images with a poonom fitted only to rib thicknesses av ^abs (p) (v _b ^abs - values are at our disposal and we examine how much a given ν'* ⁸ value differs from ν^ζμ) from a polynomial fitted to rib thicknesses, i.e. the square root of the squared error of the polynomial fitting is required.

The conditional density functions are approximated by parametric normal distributions. The approximate conditional density function for the rib thickness is:

- f V _abs (V | M _b>upper - μ) - (27) o _v ''ν2π

The approximate conditional density function can be derived in the same way for the inter-rib distance tb* ⁸ , av _b ^for the ratio of consecutive rib thicknesses, and tb' ^el for the ratio of consecutive inter-rib distances.

If the approximate conditional density function for the rib thickness ~ f V _abs (v | M^upper = μ)(18a) is determined based on the above and is available, ~ f T _a bs (t I Mb,ai _salt = μ)(19a) approximate conditional density function for inter-rib distance ~ f V _re i (v | M _b , _upper ^:::: μ)(20a) approximate conditional density function for the ratio of successive rib thicknesses ~ f T _re5 (t | Mb.aisó = μ)(21a) is an approximate conditional density function for the ratio of successive rib distances, then by using these, the most probable lower and upper rib arch positions can be found with the under-sampled rib signal ah _prC p[n], to which the corresponding g curves will give the approximate rib arches,

The task is solved by dynamic programming using the DPABPME dynamic programming-based costal arch position determination procedure, the input of which is the undersampled costal signal ah _prep [n] and the J9 DPABPME input parameters from the PBE parameter setting unit, namely ~ f V _abs (vj M _bt fei _S ó - μ) ((equation 18a)), approximate conditional density function for rib thickness, « fT^ (t | M _b>0 | _S ó = μ) ((equation 19a)), approximate conditional density function for inter-rib distance, the - f V _re i (v | Mb,feisö ~ μ) (equation (20a)), approximate conditional density function for the ratio of successive rib thicknesses, ~ f T _rs! (t | Mh,afsé ~ μ) (equation (21a)), approximate conditional density function for the ratio of successive rib distances, ν^Χμ) for rib thicknesses

47/20 fitted polynomial, ^(μ) polynomial fitted to inter-rib distances, 1¾) polynomial fitted to the ratio of consecutive inter-rib distances, v' ^d (p) polynomial fitted to the ratio of adjacent rib thicknesses, o« ^ai>s rib thickness standard deviation, ot ^abs inter-rib distance standard deviation, oThe ratio of inter-rib distances standard deviation, ov' ^e! the ratio of overlapping rib thicknesses is standard deviation. Additional elements of the J9 DPABPME input parameters are defined later. The output of the procedure is the lower and upper rib arches J10.

Simplifying the solution first, II. is described based on fig. For the sake of simple illustration, we combine the rib thickness av and the distance between ribs at 1 ~ vvt, so we only search for the upper curve of the ribs. The programming steps are defined on the μ 1 position-step distance plane, on which a point is the winter of an upper rib arch and the corresponding l step distance (vK). On this plane, the set of EKEI valid starting points (dashed range) region marks them aq _st ^. starting points, which can be the upper curve of the first rib, so the curve can be in different positions and have different steps (different rib thickness and distance between ribs are possible). In the same way, we also define the region of the set of valid endpoints of the ÉVH (dotted region), which denotes the possible endpoints q _ve8 , here are the possible upper curves of the last rib. The set of valid EPH positions belonging to the other ribs is located in the strip between the two regions, i.e. the possible rib points. In addition to the above areas, there are TZ prohibited zones (grid areas) on both sides. In positions falling into the TZ prohibited zones, the step spacing is not allowed based on the absolute limit.

After that, costs are assigned to each point. The value of the undersampled edge signal h^pfn] at position μ is considered the image cost ah _prep [μ/s], and a "probability" cost can be derived from the values of the conditional density functions obtained as input parameters at a given point.

The task is to get from the EKEI set of valid starting points region of Figure 11 to the ÉVH set of valid end points region in such a way that the cost assigned to the affected points is minimal. The possible points are in the set of EPS valid positions band width, ÉPH valid positions, that is. and I ^8b!> (μ) is the center line of the allowed band (dashed line).

/k process is described for step 3. By moving I steps along the μ position axis (indicated by quarter arcs (circles) in Figure 11, i.e. you have to move down the same amount as to the right (h), you get a horizontal line passing through the lower point of the circular arc, along which you can select the stone intersection point. The horizontal along a straight line, the estimated probability of the absolute and relative step distance must be taken into account, i.e. the step cost l ^8bs and l ^te}

47/21 relative step cost. For the latter, we add an RLVT relative step interval variation range on the former horizontal line, which contains the points that are more likely than a lower limit, and we can only step on them. The closer the relative step cost is to the center of the above range (section), i.e. the minimum of the relative step cost l^Am. This cost thus already depends on the previous step, so this cost is not assigned to a specific point, but to the step. This is illustrated by the position corresponding to step 3 in Figure 11. And the absolute step cost l ^m is smaller the closer we are to the center line of the allowed band of the step interval l“ ^bs (μ), i.e. the center line is also the set of positions with the minimum absolute step cost 1 ¹ %^. In Figure 1L, the um...? the selected vertical rib levels, and the squares represent the position of the selected rib arcs KB at the above vertical rib levels using the relative step cost l' ^e} and the absolute step cost I**.

For a detailed description of the solution, see Figure 12. In order to extend the pitch I separately to the rib thickness v and the inter-rib distance t, two horizontal axes must be taken instead of the I axis, so that both the previous rib thickness and the previous inter-rib distance can be noted and taken into account. Each Åj step is divided into two orthogonal steps, a Vj step on the v - μ rib thickness - position plane and a t, step on the t - μ rib distance - position plane. Let the point q be one of the points of the μ-νt position rib thickness - inter-rib distance space. At point q, p(q) is the value of the position, v(q) is the value of the rib thickness, and t(q) is the distance between the ribs. The point belonging to the lower curve of the i-th rib of the q^atsó, qw.fcfeö the point belonging to the upper curve of the i-th rib, μ^^ is the position of the point belonging to the upper curve of the i-th rib, the thickness belonging to the upper curve of the i-th rib , t^.fch$ is the value of the distance between the ribs belonging to the upper arch of the L th rib, where this is clear. In Figure 12, the limits must be checked and the costs are calculated in the positions marked with a light square. The positions marked with a dark square represent rib wood. The br.z the ribs, the t^. .4 current inter-rib distance, the Vb?. and < indicates current rib thickness

Some quantities must be limited to solve the task within an acceptable time. pg^ is the μ position value belonging to the N~th (last) generated curve. The boundaries of the set of valid starting points of the ÉKH containing valid starting points q^i and the set of valid ending points of ÉVH containing valid q _VS g endpoints are determined as a function of the μ position value belonging to the Nth (last) generated curve ap _g ^ on the μ position axis:

Oápfq^Jác^p^ (28)

47/22

The starting point! variable and the endpoint variable _ysg as part of the J9 DPABFME input parameters, that is. A set of valid starting points for ÉKH and a set of valid ending points for ÉVH are set. 0J3 is a good choice for both variables. To determine the limits for rib thickness v and inter-rib distance t, we need to introduce the confidence interval parameter ac, which is also part of the J9 DPABPME input parameters and specifies how many times the standard deviation of rib thickness ao _y is the permissible deviation from the polynomial fitted to rib thicknesses av^ (μ), and the number of times the standard deviation of the distance between the ribs is the permissible deviation from the polypom fitted to the distance between the ribs (μ). The polynomial fitted to the rib thicknesses v* (μ), the polynomial fitted to the inter-rib distances í* ¹ * (μ), as well as the permissible deviations defined above, i.e. co/^ and eo^, determine the effective rib thickness of the ÉVPS according to Figure 12 positions bandwidth, which is 2eo _v ^S0 \ and the ETPS valid inter-rib distance positions bandwidth, which is Scot ⁵ *, The bands defined in this way delimit the set of valid rib thickness positions of ÉVPH and the set of valid inter-rib distance positions of ÉTPH. Using the above, I can write step rules;

W- ^cc C ^te A v _bi 4·(30)

E'fov, _! sö )- ca/ ^s < U., _jys6 < v co/ ^s (31)

The ranges outside the valid positions are the TZ forbidden zones, the step rules according to equations (30) and (31) are illustrated by the light squares in Fig. 12. Equation (30) can be used for going from upper to lower arc, while equation (31) can be used for going from lower to upper. So, if the upper curve of this rib is already selected, then the lower curve can take on rib thicknesses that are allowed in the phi fcisö Fedik rib upper curve (vertical) position. In the same way, if the lower arch of the i-th rib is selected, the Hl-th rib can take on values of inter-rib distance t that are allowed in the lower arch (vertical) position of the Cover rib. So, in fact, the rib points q marked with a dark square in the figure must not fall within the range of the set of valid rib thickness positions of the allowed ÉVPH, or the set of valid inter-rib distance positions of the ÉTPH, but the ones marked with a light square, which already take up the i +1 -th rib point at the position of the Fedik rib point qj the value of the positional ¥ rib thickness or the inter-rib distance t.

The compliance with the limits is therefore always measured only in the light squares, so if you go from the lower curve to the upper one, the rib thickness does not have to correspond, only the value of the distance between the ribs t. In the same way, if we go from top to bottom, we only need to check the rib thickness v, not the inter-rib distance t, as in Fig. 12 μ - t position - inter-rib

It is also clearly visible in the 47/23 distance coordinate system, many dark squares slide into the TZ prohibited zone, while the light squares are all the way in the set of valid intercostal distance positions of the ÉTPH.

In addition to absolute limits, we also use relative limits, which depend not only on the current μ position, but also on the previous rib point q. Figure 13 illustrates the process of moving from an upper rib arch to a lower rib arch on the μ - v position - rib thickness plane, where ab; the transition from the upper curve of the i-th rib to the lower curve of the i-th rib b< is visible. If we move from a bi, _!e í _S ö i-th rib upper arch (PHfeisö position) to the i-th rib lower arch (μ^,^ position), then the rib thickness corresponding to the i-th rib lower arch must be as follows exist:

“ v _bi .. _J>a!só crt^ < V _{bi 3jsó} - v^v' (p _bi .. _1;31só ) < v _bl _ _Wsó c< ^ef (32)

But since the values of the rib thickness av do not change when walking on the μ -1 position - inter-rib distance plane, i.e. ^V bí _s íe5sö ~ ^V bj-l.aísó (-''), so the limits can also be expressed using the lower arc of the i-th rib q _b :

“ ^bi.íéisö^^v ~ ''bi.afsó ~ bi,feis6(úbi,upper) — bCupper^v 04)

In Figures 12 and 13, the BWT rib thickness variation range based on equation (34) is 2v _b i _(f a _sö co _v ^{rel )} , and the mean value of the BVVTK rib thickness variation range is Vbi,feisőY ( dbi.fciső) ·

The process of moving from the lower rib arch to the upper rib arch is illustrated in figure 14, where the transition from the lower arch of the i-th rib (gbraisó position) to the upper arch of the i+l rib (gbHj. feisö position) of bt is shown, the position μ -1 - intercostal distance on the plane. If we move from the lower arch of the i-th rib of a bj.ajso to bj+jje it is. í+lth rib to upper arch, then ibi+j. and the inter-rib distance belonging to the upper arch of the i+lth rib must be the following:

Top ~ Top ~ Bottom ~ Bottom (top)

But since when walking on the μ - v position-rib thickness plane, the inter-rib distance at values are copied f ™ t ^L bi.öJ$á Hú,upper, so the limits can also be expressed using the upper curve of the ith rib qbi.febő:

Elower ^C<7 t “ Πί+Ι,upper Hi.alscd (Pbi,abó) ^bi,iW ^CO t (37)

In Figures 12 and 14, the BTVT inter-rib distance variation range based on equation (37)

2 tbi. isóCO _t ^Íeí , while BTVTK is _the mean value of the variation range of the intercostal distance

The choice of confidence interval c (part of 19 DPABPME input parameters) requires careful consideration, which affects both the process and the result. If we choose a large one, the process slows down a lot, if we choose it too small, the search may discard possible solutions. Based on the tests, the confidence interval c:! .choice 7 is a good compromise solution. Then the procedure accepts 90% of the sample set. This does not necessarily mean that the segmentation of 10% of the ribs on the model foam will be incorrect, but only that 10% will not be accurate, i.e. it will shift the accurate rib arch to a position that already falls within the permissible range.

Looking at Figure 12, it can be observed that the upper point of the first rib aq _b is completely missing from the figure, and the lower point of the first rib has no projection on the μ -1 position-interrib distance plane (the bright point is not a bearing point q).. This is because , because these points are special, since they have a position μ, but they are projected onto the μ - v position-rib thickness and / or μ - t position inter-rib distance is not defined on the plane

The starting points are upper rib arches for which the equation (28) is fulfilled, but the values of the rib thickness v and the inter-rib distance t are not determined, since there were no other rib arches before them, from which thickness or distance could be calculated, and (34) and (34) and relative limits according to equations (37), but not even absolute limits according to equations (30) and (31). So these points are practically one-dimensional, they are characterized only by the position μ.

Moving on from these points, we can choose any rib thickness that corresponds to the limit according to equation (30), the relative limits are not valid and cannot be used at this step. The lower points of the first rib q^^ thus already have the rib thickness value v, but still not the inter-rib distance t, this too. means that the value of the inter-rib distance t is only subject to the absolute limit according to equation (31). The q«. The upper points of the 2nd rib already have the inter-rib distance value t, so they are full-value q rib points.

In order to be able to choose from among the possible new points, we assign costs to each step. Certain costs depend only on the position μ of the new point, others on the new point μ -1 position · rib distance, or the μ · v position-rib thickness projection, but there are also costs that depend on the starting point of the step. The cost function consists of three main parts:

- k^ image cost

- k _p probability cost

- k _s rib thickness cost

The total cost:

47/25 (38) where aw, _m image cost weighting factor, w _p probability cost weighting factor, w _s edge density cost weighting factor. The above costs and weighting factors are part of the J9 DPABPME input parameters,

A ki _m image cost aq _; Moving from the i~th edge point to the iHth edge point, the value of the down sampled edge signal ah _w [n] corresponding to the current μ position of the edge point qn í:

(39) if q _w is an upper arc if q _M is a lower arc (40)

The two-state variable Ψ has a value of 1 if an upper rib arch is selected, and -1 if a lower rib arch is selected, and is the offset of the fitting circular arc. We would also like to assign a cost to each step that expresses the statistical probability of the given step. The probability cost k _p is defined so that the cost of the entire road increases monotonically as the probability of the rib configuration belonging to the road decreases. Accordingly, if there is a rib line defined by a series of rib points qm qp, ... and another rib line defined by a series of rib points qp, qp> .... q^, then the path connecting the rib points qp, qp, -·> %» and between the probability and the cost of the path connecting the rib points qp, qp, .... q»» is:

Hq _ib qp,..., q^) > P(q _u qp,..., q _k )

Way-; Qm.) kp(qjn qp...., q^) (M) relation must be satisfied. The probability of one or more configurations should be the transition probabilities, i.e. the product of the probabilities of each step, in general:

P(q _i ,q _2! ... _; qj^P(q _i ->q,)P(q ₂ ->q ₅ )...1^,. ->q _n ) (42)

Since we want to find the optimal solution by searching for the path with the lowest cost, we need a cost that is not a product of partial costs, but a sum of partial costs, i.e. along a path consisting of a total of N points, the probability cost kp is

N

->q _w > (43)

Correlation can be used. The latter is possible if the cost is defined as a negative logarithmic probability, i.e

M qi, 02- ₅ q«) ^:::: · -^lnP(q; -» q _i+I ) (44) te te te

47/26

Nl ⁼ - ^1η Π ^ρ ^*^) < ⁴⁵ ) --InP^q,....^) (46)

The transition probability consists of four components:

p/ _n . " ... (qi "* q _w )P _v ' ^e! (Qi ”» q _i+! ) if q _{ an upper arc '14; 4í+J ] _n ab5 _z - í , . .

IN (q, ->q _i+ :)l ⁵ _t (q; -» q ₁₊ J if q _; an also arc

The individual components: Pv^fq. qm) ^ρ ( ^ν * ^{3 v} (4m) 1 μ = q(q.)) (48) P _t ^aos (q. -> qi+I) ⁼ P(t ^abs t(q^t) i μ - ü (q _} )) (49) / . V P.'Vq, q _w ) - P| v" ^! = Vfe. | _μ = _μ( ,.) ( ₅₀₎ kv(qj J ρΛο, -»q _w ) - p| t· - τΗ I > ^l “ < ⁵¹ > kf(q,) /

The transition probabilities according to individual equations (48)...(51) are given by the conditional density functions according to equations (18)...(21). We approximate these with normal distributions. In the end, the probability cost function for rib thickness ak _pV abs av ^abs :

of N

.... yU * | ^(Aím)·” fi 2o/- ^sZ = (N -1 )taN“ £ feat w 2o _v ^a “ ⁵ = (Nl)z + ^

(52)

In equation (52), we denote the expression constant ln(o _v ^abs k(2^) by z~. Using the above approach, we obtain a cost function for rib thickness av^ which sums up the squared deviations measured from the conditional expected value of the thickness, and a step number of the path for the quantities according to equations (49)...(51) we also get cost expressions structured according to equation (52), so the more probable a path,

47/27 is closer to the expected values and gets a lower cost. Omitting the term containing the constant and the multiplier of two in the denominator from equation (52), since they do not affect the comparison of the probabilities of different paths, the final, simplified form ak _pv abs for the probability cost function for rib thickness:

(«) m σ.

And the specific cost functions for one transition based on (48).. .(51):

<A -* q _M ) =(54) σ _ν

V. _w (q, q„.) =($5) σ, ' / ·\ ^{2 k} „™n -* qw ) '--5<

^σ ν

¢.0 “ —a<

σ. t

The kpj probability cost (function) for a transition based on (47):

. , . + if an upper arc k ,(q—> q. ,) - < Ö8) ” ' ^{, J} + if cp is a lower arc. According to equations (39), (40), the image cost kj _ra can be negative and , that the probability cost ak _p cannot compensate for this, so it may happen that as the searcher moves forward along a path, the cost will become smaller and smaller, so that longer paths will be cheaper than short ones, this can lead to the generation of incorrect rib arcs. To compensate for this, we introduced _the rib density cost. The higher this is set, the fewer rib arches the search results in. The choice of this largely depends on the parameterization of the under-sampled rib signal ah _prep [n]. was the pre-processing. Rib density k _s cost: a value of 2 is usually a good setting, but this parameter definitely requires tuning. It is worth weighting the probability cost kp to offset the cost of the image. The probability cost weighting factor w _P : a value of 0.7 can be used, but it requires tuning.

47/28

Oi As At At os on it

Using the above, the first step of the implementation is the selection of individual bands and ranges on the μ - v position-rib thickness and μ -1 position-rib distance planes (Figure 12 and Figures 13 and 14);

• designation of the set of valid rib thickness positions of the ÉVPH (allowed band) on the μ - v position-rib thickness plane, the positions are located in the band defined by the 2co _v ^801> double confidence interval-times rib thickness standard deviation, which defines the bandwidth of the EVPS valid rib thickness positions, whose center line is av ^8te '(u) poünom fitted to rib thicknesses, ♦ designation of the set of valid inter-rib distance positions (allowed band) of the ÉTPH on the μ ~ t position-inter-rib distance plane, the positions are the 2ca ^abs double confidence interval-multiplier inter-rib distance that gives the bandwidth of the ETPS valid inter-rib distance positions they are located in a band bounded by the distance standard deviation, the center line of which is the polynomial fitted to the inter-rib distances ^(μ), • designation of the set of valid starting points of the EKH on the μ - v position-rib thickness plane, ♦ designation of the set of valid endpoints of the ÉVH on the μ ~ v position-rib thickness plane , * selection of the set of valid starting points of the ΕKH μ - t on the position-interrib distance plane, * selection of the set of valid endpoints of the ÉVH on the μ -1 position-interrib distance plane, ♦ designation of the TZ prohibited zones on the μ - v position-rib thickness plane and μ - t position interrib distance on the plane.

After selecting the domains, the successive steps are based on the following rules:

« we use light and dark squares for the steps, in the light square the compliance with the limits is checked and the probabilities are calculated, the dark squares are the possible rib points from which the next step is made, • on the μ - v position-rib thickness plane, the μ position direction steps starting from the light squares (quarter circles) mark the location of the lower rib arches belonging to the given upper rib arches.

47/29 * on the μ t position-intercostal distance plane, the steps in the direction of μ position starting from the light squares (quarter circles) mark the location of the upper costal arches that touch the respective lower costal arches, ♦ a step starting from a light square always ends in a dark square, ♦ performed on any plane light square after a dark square step, you have to step on the other plane with an unchanged μ position, dark square .....* with a dark square step. If the dark square was created on the μ - t position-rib distance plane after a light square —» dark square step, then at the given μ position value, by moving (horizontally) onto the μ - v position-rib thickness plane, the new dark square is created, which The position of rib thickness v is the same as the original value, one level higher, i.e. the coordinate of the new dark square v rib thickness HL t inter-rib distance in the opposite field is determined by the following rules:

• the rib thickness values v do not change when walking on the μ - t position-rib distance plane.

* when walking on the μ - v position-rib thickness plane, the inter-rib distance t values do not change.

♦ In the environment of the new dark square created after the dark square -» dark square step, the value of the rib thickness av on the position-rib thickness plane μ - v, and the value of the inter-rib distance at on the position-rib distance plane μ ~ t can be varied in a given range, with a constant position μ , so a new light point can be created, with which the procedure can be continued, ♦ the step is only valid if the created new light point remains within the set of valid rib thickness positions EVPH and the set of valid inter-rib distance positions ΕΓΡΗ, ♦ a dark point can avoid the TZ also in the forbidden zone, ♦ when starting the search, we select a light point with the parameters μ-q rib position, v _q rib thickness, t^ intercostal thickness from the set of valid starting points of the ÉKH, (qs; upper rib arch of the next first rib).

Equations (28) and (29) apply to the selection of the set of valid starting points of the ÉKH and the set of valid end points of the EVH.

When moving from the upper curve to the lower curve (Ί Fig. 3):

lake

47/30 • based on equation (34), the mean value of the BVVTK rib thickness variation range AíiJehAb (bhi, upper)?

♦ based on equation (34), the BVVT is the rib thickness variation range o _w5 , ♦ the dark square is the dark square resulting from the μ -t position-inter-rib distance plane, ♦ the light square can be freely selected in the variation range for further progress.

When moving from the lower arch to the upper arch (Fig. 14):

* based on equation (37), BTVTK is the mean value of the inter-rib distance variation range « based on equation (37) BTVT is the variation range of inter-rib distance Sth^ai^c o^, « the dark square is the dark one resulting from the μ - v position-rib thickness plane square, ♦ the light square in the variation range can be freely selected to proceed.

The costs assigned to individual steps and their components are given by equations (38)...(40), (47) and equations (54)...(57)

Based on the above, the search is carried out using the Dijkstra algorithm (the least-cost path-finding algorithm developed for graphs). Both the discrete rib arc search start points satisfying equation (28) and the discrete rib arc search end points q _w satisfying equation (29) are considered graph nodes. After that, we pick a starting node, preferably above the region of the set of valid starting points of the EKH according to Figure 12, containing the possible starting points q^. We take q possible edge points as graph nodes, which correspond to the constraints according to equations (30), (31), (34), (37). The starting node and starting points (graph) are connected by edges. We assign costs according to equation (38) to all further edges between pairs of possible nodes, and then a search must be started from the starting node to the discrete endpoints defined by equation (29) on the graph. The q points belonging to the path with the least cost give the resulting rib arches. Since it is not stipulated that the last point along the path must be a lower arc, if the path ends with an upper arc, it must be discarded. The output of the process is the II0 lower and upper rib arches, and the .111 lower and upper rib arches~2, and the .112 lower and upper rib arches-1 signal with the same content as the JW lower and upper rib arches passing through the V selector.

The previously defined rib arches must be specified. One possibility for this is the DPACSPE dynamic programming-based bone contour refinement procedure connected to one of the outputs of the V selector, which looks for curves similar to the J12 lower and upper rib arches- 1 as starting curves in the environment of the originals, and selects the one with the intensity on both sides

47/31 difference is the maximum. To explain the essence of the procedure, Figure 15 shows the initial curve ag and a possible modified (result) curve g'. In the figure, the circles correspond to one point of the x-ray machine, the arrows illustrate the possible further steps 1 from a given point, limiting the different options to three. For the sake of clarity, the further options are not marked for all points. The procedure is described in more detail in the 16 , is described using Fig. The input of the procedure is a curve, in this case curve ag, given by the series of points Pi (i ~ l...n) (the "vertical" coordinate of the points relative to curve g is 0, so it is not marked separately), the procedure requires the J13 DPACSPE input parameters, such as the distance of the (division) points δ, the value range of the shift parameter m, k _c step cost, kcs the cost of moving to the side and the (image) intensity values in points I(pi.j) ap,j. The direct output of the procedure is a series of p μ points (yector) defining a curve modified into a curve ag*.The steps of the solution:

• the initial curve g is divided into n parts, pj is the i-th point of the initial curve (i - 1,. .n), • at every i-th point of the initial curve g, lines perpendicular to the tangent of the curve at the given point are set (e.g. in Figure 16, p , in the case of a point, the line passing through the points a and Pj.), • points (j = -m.. .0., ,+m) are taken with an offset on the perpendicular lines, * the pij point is p, point j-th offset point (e.g. in figure 16, point p, +mth shifts pp-m), • 1( Pm) is the value of the (image) intensity at point pij, ♦ previous intensity values are copied into an intensity array T(pjj), ♦ the elements of the intensity array T(ptj) are normalized, then the result is copied into the normalized intensity array T(ptj), the elements of the latter array are the normalized intensity values Pipy), « the values of the normalized intensity array TApm) are modified based on a given rule system, filtered, * two types we perform filtering, filtering in the normal direction and in the tangential direction, • when filtering in the non-normal direction, the new Γ*(ρ^) belonging to any pij point is inserted in the normal direction t, normalized intensity value is obtained by a weighted linear combination of the original E(pij) normalized intensity values in the vicinity of the point. Figures 17a, 17b, 17c show one possible weighting, with the weighting numbers written in the squares, with the weighting values corresponding to the environmental point in the normal direction taken into account in increasing numbers. The figures take into account points 1-1, 2-2, and 3-3 above and below the starting point, respectively. For example, the 17c. Fig. is the starting point

Moving away from point 47/32 upwards, in a row -4, -2, -1, while moving away from the starting point downwards, in a row it uses 4, 2, 1 intensity value weighting factors.

• the resulting filtered values are copied into the filtered, normalized, intensity array T'*(pij), • tangential filtering is then performed on the elements of the filtered, normalized, intensity array T'*(pij). The results are stored in the twice-filtered, normalized, intensity array T'**(Py), • the order of the filters (normal direction and tangential direction) can be changed, • a weighted directed graph is created from the p points;

Mpmj Pij.j) ^::: Τ(ρ _ίχΗ ) + k _cs (59) k<(Pi-ij -»Pij) ^::: Τφψ(60)

Mpi-ij Ρφΐ) = Τ(ρ _φ! ) + k _cs (61) ak _c with step cost, ak _cs with side step cost and Ήρ^) array containing image (point) intensities as image cost.

♦ we take a point (pj) as the starting point, the search starts from here k _c (pi -» poj) = T(Poj) (j =~m...0...+m),(62) * we take a p _n points where the number of points on the curve na k _c (Pn-Q Ρη,ο) 0 0 = ~m...0...+m),(63) • the output of the procedure is p. _} , from point ap _T , is given by the least-cost path. The search is possible moving forward by position, because if all kg accumulated step costs are available up to the (il)th position, then the accumulated step cost ak _c (pij) can be calculated in the i~th position as well:

±c(Pí,j) min (kc(pi-].k)+k _c (pj.j. _k --> c(pi.i)) (64) * the process can also be parallelized, since it. the cost can be calculated independently in a given position, • the optimal path can be obtained by selecting the one with the lowest accumulated cost from among the points belonging to the last position and moving backwards along the neighbors with the lowest cost.

The procedure was performed with VBI definitive rib cages.

The above method can be extended to specify the lower and upper rib contours together according to Figure 18. In this case, the two rib contours can be prescribed to move somewhat together, which corresponds to the fact that the lower and upper rib contours are more or less parallel. With this method, the search procedure can be made more stable, such things can be excluded

47/33 solutions when the two curves, two contours of the same rib cross each other or are very far apart. The DPA2CSPE dynamic programming-based double bone contour refinement method, which implements the above and is connected to the other output of the V selector, whose task is to search for curves in their surroundings similar to the lower and upper ribs of J11 given by a series of points during matching. Among the curves, it selects the ones with the maximum difference in intensity on both sides, taking into account the parity of the resulting curves during the search. The input of the procedure is a pair of curves, ag^ which gives the initial lower curve and gw the initial upper curve p^ , the i»th point of the initial lower curve, ap^ ; the i-th point of the initial upper curve is a series of points (i™ l...n), and the input parameters required for the procedure are the distance of the dividing points δ, the value range of the displacement parameter m, ak _c step cost, ak _CS i _Jna ^g smoothness cost _{. _} the i-th point of the p*^ modified lower curve giving the lower curve and the g'jwa modified upper curve, the i-th point of the modified upper curves ρ&Μ,ί is a sequence of points/vector) (i™ 1. < m). Solution steps:

• the initial lower curve g^a and the initial upper curve gfcus are divided into n parts, the i-th point of the initial lower curve í, the i-th point of the initial upper curves pr^ , (í™ 1 ···<

» at every i-th point of the starting curves, we set perpendicular lines to the tangents of the curves, ♦ we take points (j - -m...0... +m) with an offset on the perpendicular lines, • ap^jij where the starting upper curve i -th point is the jth shifted point, ap^ya is the jth shifted point of the i~th point of the starting lower curve, ♦ I(py) is the value of the (image tension) in the upper y points, l(p^ _é ap _afe6 jj points is the (image correction value, • the above intensity values are copied into the upper curve intensity array T(pw; y) and the lower curve intensity array T/p^ y) and the elements of the ay) upper curve intensity array and the Typsis<>lower curve intensity array are normalized, then the result is copied into the TXpfeíssi.i) upper curve normalized intensity array and Ί vp _aÍSÓ y) lower curve normalized intensity array, the elements of the latter arrays are ar/PíUsó y) upper curve normalized intensity values and y) lower curve normalized intensity values.

47/34 • the Ί'Χρ&ΐ»; ü) the values of the upper curve normalized intensity array and the lower curve nonnormalized intensity array T'(p _a a<> y) are modified and filtered based on a given set of rules • two types of filtering are performed, normal and tangential filtering * when filtering in the normal direction, any pfejs.ö The new F^íp^sa y) curve belonging to the starting upper curve point y is the filtered, normalized intensity value in the normal direction of the original one located in the vicinity of the point. The normals of the upper curve I'fp&fes í.j) are obtained by the weighted linear combination of intensity values according to Figure 17, the resulting hair values are stored in the filtered, normalized, intensity array of the upper curve T'NpmAij), ♦ when filtering in the normal direction, any initial p^, q The new ΓΉρ^ μ) normalized Intensity value of the lower curve point, normalized in the lower curve point, is obtained by the weighted linear combination of the normalized intensity values of the original FCpaj^ ij) lower curve in the vicinity of the point according to Figure 17, the resulting spiked values are T'*( lWq) is stored in the filtered, normalized, intensity array of the lower curve, • we also perform tangential filtering on the elements of the filtered, normalized, intensity array of the upper curve Γ*(ρ^ y). The results are stored in the T'**(pr _e u§ j ₀ ) upper curve pierced twice, normalized, intensity array.

• the T ^! *(p _a uc rí) lower curve filtered, normalized, intensity array element, after that we also perform tangential filtering. .The results are stored in au) lower curve twice filtered, normalized, intensity array.

♦ the order of the filters (normal direction and tangential direction) can be changed.

The search for the optimal path is performed simultaneously on the starting lower curve and on the starting upper curve g&us according to Figure 18. The points of the two curves can be mutually matched, so a 1 position includes two offset values sj, the offset of the upper curve, and the offset of the lower curve. For this reason, the array on which the search is run expands to three dimensions, according to Figure 19. The slice of the three-dimensional array corresponding to a given position i is shown in Figure 20 π**? in case of possible offset and possible difference. During the optimum search, the working points can be located in the fields without crosses according to Figure 20, the boundaries of which define the permissible deviation between the lower and upper rib arches. The following conditions must be met for each point of the road, for shifting the i-th point of the upper curve sfepi) and for shifting the i-th point of the lower curve s?(pi):

i (s,(p _; )~ Mp^July 1 (66)

47/35

So, steps are allowed in which we step to the side at most once along any curve, but we can step on both at the same time.

We assign costs to the steps. The k _c step cost, the smallness is smoothness! consists of cost and kckéö image cost;

MPm -> Pi) ^:::: (p _w -> Pij + k^íPi) (67)

The smallness smoothness cost consists of the cost of moving to the side ak _<s and the cost of moving two curves apart ak _C d:

^k c» (Pi-ι P. W _cs (p _w -> _Pj ) + ( _Pi _, -> pi) (68)

The cost of moving to the side kos and ak _c the cost of the step that moves the two curves apart is written down as _the Sávo average shift ((sj+ S2)/2), ad is the average step difference d belonging to the step difference (Figure 18), ay _s is the weighting factor for moving to the side and ay _d using a weighting factor that moves two curves apart:

(Pí_ _} Pi) y _s I savo (Pi) ” s _AVG (P> - ¹ ) I (69)

MPí-1 Pi) = Yd I ^d Avo(P.)- ^d Av _O (Pi -1) I (70)

It is advisable to set the weighting factor for moving to the side y _s and the weighting factor for moving the two curves apart in the interval 0...1, since the image cost of the image _C can also change in this interval. For both weighting factors, 0.2 is a well-suited value in practice

The image cost is generated from the pre-processed T(pf _e [ _s6 and T(p _a i _s6 ij) intensity arrays of the two curves:

k _ctop (P,) = f (T í p _te<; , J- T (p _a . _sób p(71)

The function f can be a simple addition, but it can also be a more complex and sophisticated function, e.g.:

i /(^)==1-(1-^,)(^(72)

After the normalization of the image cost, the path with the least cost can be searched moving forward by position. ^(p,) is the accumulated step cost for one cell of the search array:

MP, ) ^::: _ min Mt) + ^k e (Pi)(73) where P _SK>ms (p/) denotes the points p.., determined by equations (65), (66). During the route search, the cost of the lowest-cost route to the given point must also be stored here. The optimal path can be obtained if we select the one with the lowest accumulated cost among the points belonging to the last position and move backwards along the neighbors with the lowest cost. The procedure was performed with VBI definitive rib cages.

Claims (4)

47/36 Advantages of the method according to the invention: • using a guide field, in contrast to existing methods, the rib arches are not determined analytically (with a function or a combination of functions), but with series of points, therefore the results obtained are more accurate and better fit the real rib arches, • the procedure can handle up to 8-11 ribs, in contrast to with known methods that deal with a maximum of 5-7 ribs, • the known methods only segment within the outline of the lungs, while the proposed method can generate and continue the rib lines beyond that, ♦ the used method determines the shape and position of the ribs independently of each other first, it calculates the direction field that determines the shape of the rib arches, from which it generates possible rib arches starting from the center line of the Cmid ribs, and then the selection of the most probable rib positions follows, • the former separation is advantageous because the dimensional subproblems during the solution of the problem are significantly smaller, • the rib outline refining process at the same time ke zeli the lower and upper arch of each rib, since the approximate parallelism of the lower and upper arch of each rib can usually be assumed, the application of this rule results in a large number of incorrect segmentations! excludes the possibility, • the method gives more accurate results than the known methods, because on the one hand, the rib lines generated by the method fit better on the original ribs, and on the other hand, the generation of incorrect rib lines is significantly smaller compared to the known solutions. PATENT CLAIMS 1. A procedure for analyzing rib shadows and determining the contour of the shadows on chest radiographs, which processes a double-sided lung radiograph (KTRF) by sequentially using a preprocessing unit (EFE), an image coordinate system adjustment unit (KKOB), the double-sided radiograph it uses a one-sided converting unit (KOEO), a unit for setting the image frame (KKEB), a pre-segmentation unit (ESZE) that determines the lung outline at (q), followed by an evaluation unit (KE), characterized by 1···· SZTNHd 00200983 47/ that the evaluation unit (KE), ^the image pre-filter unit (ESZÜ) connected in series to the edge search, the unit that searches for convex circular edges in the image (KKAÉKX, then the direction field matching unit (H), then the one-dimensional edge signal generating unit (EDBE ), this is essentially a dynamic programming-based costal arch position determination procedure (DPABPME), followed by a selector (V), which, depending on its position, either a dynamic programming-based bone contour refinement procedure (DPACSPE) or a dynamic programming-based double bone contour refinement procedure (DPA2CSPE), and separate parameter setting units (PBE) necessary for the operation of the above units are used, where the image pre-filter unit (ESZÜ) performs anatomical noise filtering on the pre-segmented image (.11) connected to its input, with the ESZÜ input parameters (12) coming from the parameter setting unit (PBE) and performs edge extraction, the output is the pre-filtered image (J3), where the convex circular edge search unit (KKAÉK) works as k that starting from the pre-filtered image (J3). as input data, the parameters of circular segments of matching circular arc length (L) intersected in the middle by a short vertical segment passing through pixel (p), namely matching circular arc curvature (k), matching circular arc radius (R), matching circular arc inclination angle (a), matching circular arc vertical shift (s), it traverses the space between the pre-fixed extreme values of the parameter with a given resolution, until the difference of the intensity values summed at the edge of the band (dk) of a given width above the circular arc and below the circular arc reaches its maximum value in the case of the lower arc, in the examined pixel (p) the the first approximation value of the direction field is given by the direction of the tangent to the circular arc in radians. when searching for the upper edge of the ribs, a minimum search must be used, the parameter ranges to be tested and the resolutions to be used (14), which come from the parameter setting unit (PBE), are as -0.45 <a < 1.15 , | H _e | = 25, 2< k <8, |H _r | - 4-0.025 < s < 0.025, | HJ - U,n =20, k = 0.075, d _k = 0.0015, the output of the unit is the matched directional field (15), where the directional field matching unit (II) works so that starting from the unmatched directional field (J5), as from the input signal, the parameter setting unit (PBE) adjusts the direction field to the obtained inclination angles by repeatedly applying the procedure according to the convex circular edge search unit (KKAÉK) in the vicinity of the points (16) of the hexagonal grid designated within the outline of the lung, which is defined by the direction field function (Φ(ρ)) is described, which gives the angle of inclination of the possible rib arch at any point of the lung and is also the output of the unit, where the one-dimensional rib signal generating unit (EDBE) works in such a way that the direction field function (Φ(ρ)) as by using an input signal, it determines the intersection points of the assumed bearing arches with the midline of the ribs (cmúO, then the assumed rib lines 47/38 points (q), moving in two directions with step intervals (u) by applying the formulas q-+i = φ + u(cosaj sina) and qn - u(cosa, sina), then for the ith assumed rib line the pixel intensity function (Iimage(p)) and the set of generated points of the i-th curve (G,) using ah _; using the formula creates the intensity assigned to the ith curve, (hj), then ah[n] = h _n +i if 0 < n < N, otherwise using the expression 0 it determines the edge signal (hjn]), then the edge signal (h[ nj) smoothes, differentiates the smoothed ridge signal (hsimitM), normalizes the differentiated ridge signal (hdiffn]), downsamples the normalized ridge signal (h _nonn [n]), gives the undersampled ridge signal (hpr _ep [n]), as the unit output signal, where the dynamic programming-based rib position determination procedure (DPABPME) which determines the positions of the lower and upper rib arches, describes the operation of the algorithm, whose input is the undersampled rib signal (h _pfe p[n]) _and whose input is the DPABPME coming from the parameter setting unit (PBE) input parameters (J9), such as the approximate conditional density function for the rib thickness ( V _abs (v | M _b , _fe j _sö = μ)), the approximate conditional density function for the inter-rib distance ( ~f T _a b _S (t | M _ba | _Sb - μ)), the successive rib thickness approximate conditional density function for the ratio of ok ( ~f V _rcJ (vj M _b> &í _S ö - μ)), approximate conditional density function for the ratio of successive rib distances ( ~f T _re i (t | “ μ)), the polynomial fitted to the rib thicknesses (v ^abs (p)), the polynomial fitted to the inter-rib distances (t ^ab \g)), the ratio of the successive inter-rib distances, the fitted polynomial (t^'íp)), the successive fitted to the ratio of rib thicknesses, polynomial (ν^Χμ)), the rib thickness standard deviation (ov ^abs ), the inter-rib distance standard deviation (ot ^abs ), the ratio of successive inter-rib distances standard deviation (σ/ ^δί ), the standard deviation of the ratio of consecutive rib thicknesses (oy ^,e! ), the confidence interval (c), the image cost (km), the probability cost (kp), the rib density cost (ks), the image cost weighting factor (Wira), the probability cost weighting factor (wp), the rib thickness is the cost weighting factor (ws), the starting point variable (estartλ ^{is the} end point variable (ενί&), the output is the lower and upper rib arcs (J10), the first step of the procedure is the selection of individual bands and ranges, the position rib thickness (μ -- v ) and the position-to-rib distance (μ -t) on planes: • selection of the set of valid rib thickness positions (EVPH) on the position rib thickness (μ - v) plane, the positions are located in the band defined by the double confidence interval times the rib thickness standard deviation (2co _v ^aby ), which gives the bandwidth of the valid rib thickness positions (EVPS), whose center line is polynomial fitted to rib thicknesses (v ^abs (g)), 47/39 ♦ designation of the set of valid intercostal distance positions (ÉTPH) on the positional intercostal distance (μ -1) plane, the positions were delimited by the double confidence interval times the intercostal distance standard deviation (Sco^) giving the bandwidth of the valid intercostal distance positions (ETPS) they are located in a strip, the center line of which is pollnom (t ^afes (p)), • selection of the set of valid starting points (EKH) on the position-rib thickness (μ - v) plane, • selection of the set of valid end points (EV H) on the position-rib thickness (μ - v) on the plane, ♦ selection of the set of valid starting points (EKH) on the position-inter-rib distance (μ -1) on the plane, * selection of the set of valid end points (ÉVH) the position-inter-rib distance (μ -1 ) plane, » the designation of the prohibited zones (TZ) on the position-rib thickness (μ - v) plane and the position inter-rib distance (μ - t) plane, which are on both sides of the set of valid rib thickness positions (ÉVPH), and the valid inter-rib distance pos are located on two sides of the set of itions (ETPH), the second step of the procedure, after the selection of the ranges, the steps within and between the ranges, which are carried out based on the following rules: • we use light and dark squares for the steps, in the light square the compliance with the limits is checked and the probabilities are calculated, the dark squares are possible rib points from which the next step is made, ♦ on the position-rib thickness (u - v) plane, the position starting from the light square Marking the location of the lower rib arches belonging to the given upper rib arches with steps in the (μ) direction, starting from a light square as the center with quarter circular arcs with a radius of possible rib thickness, where the lower endpoint of the circular arc is the possible lower rib arch, • the position-intercostal distance (μ - t) on the plane position starting from a clear square with steps in the direction of (μ) determining the location of the possible next upper rib arches following the given lower rib arches, possible intercostal distance starting from the light square as the center with quarter circular arcs with radius, where the lower endpoint of the circular arc is the possible next upper rib arch, • some clear a move starting from a square always ends in a dark square it's happening. 47/40 • after a light square —» dark square step performed on any plane, you must step on the other plane with an unchanged position (μ), with a dark square —» dark square step, if the dark square is on the position-intercostal distance (u - t) plane was created after step light square > dark square, then at the given position (μ) value, moving horizontally to the position-rib thickness (μ - v) plane, the new dark square is created, whose rib thickness position is arc) matches the original value, one level higher and, i.e., the new dark square rib thickness (v) or the intercostal distance (t) coordinate on the other plane is determined by the following rules: • the rib thickness (v) values do not change when walking on the position-rib thickness (μ -1) plane, • the rib position (t) values do not change when walking on the position-rib thickness (μ - v) plane, • the dark square —» in the environment of a new dark square created after a dark square step, the value of the rib thickness (v) on the position-rib thickness (μ - v) plane, and the value of the inter-rib distance (t) on the position-rib distance (μ - t) plane in a given range , can be varied with a constant position (μ), so that a new light square can be created, with which the procedure can be continued, » the step is only valid if the created new light square is on the set of valid rib thickness positions (ÉVPH), or remains within the set of valid intercostal distance positions (ÉTPH), • a dark point can also enter the prohibited fields (TZ), the third step of the procedure is the possible set of valid start points (q _;;iart ) and valid end points (q _w ). selection of » something is. ( 1< p(q _¥t& ) < μ,, based on equations, using the starting point! variable (a^n), the ending point variable (Svag) with a value of 0.13, and then the search in the set of valid starting points (ΈΚΗ) every rib position (μ _Λ ), rib thickness (v _q ) distance between ribs starts from a light square with parameter CQ, as the first possible rib upper rib arch (qhi.fosö), when moving from the upper arch to the lower arch: ♦ based on the mean value of the rib thickness variation range (BVVTK) ^:::: v^gfes^Vb^VobiJss;!), ♦ based on the rib thickness variation range (BVVT) ™, ♦ the dark square is the position-interrib distance (μ -1) a dark square arising from a plane, 47/41 • the light square, in the variation range, can be freely selected for further progress, when moving from the lower arch to the upper arch: * based on the mean value of the inter-rib distance variation range (BTVTK) ~, • based on the inter-rib distance variation range (BTVT) ^::: σ^ι, * the dark square is the dark square resulting from the position-rib thickness (μ - v) plane, • the light square in the variation range can be chosen freely for further progress, as the fourth step of the procedure, a cost (k) is assigned to each step, according to the equation ak ~ w _jjs k _jíö -r fw _s k, where the weighting factor (w^) is the image quality. probability cost weighting factor (w _p ), edge density cost weighting factor (w _s ), the image cost (kim) moving from the i~th edge point (q,) to the Hl~th edge point (qm) of the undersampled edge signal (h _?fi 4nj) the Mth. for the current position (μ) of rib point (qin), where corresponding value MAí the probability cost a ^€ hu ) ⁺ Ok “7- q,, ₅ j (qi “* Mi a ) * t««íi Oh 1 ha q^. an upper arc if q^, a lower arc for transition (k _P 0 and if q. egv upper arc where if q, egv lower arc is the probability cost for one step in the case of rib thickness (k _pva t>sO* in the case of inter-rib distance for one step relevant probability cost (k^O, (q, -> q _b ,) ® , _{in the case of} relative rib thickness, probability cost for one step (k^en), k........Jq, ~ -----------------i------------------a relative in case of inter-rib distance, probability cost for one step ⁵ ...... ^T (k _pteS ) ₅ k^fq, -^q^ ) -*0--^---------i-and finally the set value of the rib density cost cn (k _s ) is 2, as the fifth and final step of the procedure, a search is started using the above costs Using Dijkstra's algorithm, both Oáuíq^..,)^^ bounded discrete SALT 47/42 rib arc search starting points (q _sm )> «imd a discrete rib arc search bounded by the equation (1™s _m )g <p(q _m )áp considering end points (q _ve g) as graph nodes, and then taking the rib arc as the starting node set of search starting points (q _s w*), i.e. the set of valid starting points (EKH) over the region, as well as including additional possible edge points (q) such graph nodes that also correspond to a y.bs ab* »· x , y.bs as a limit, &QJ:· x ... . - ce, s ~ t (μ _ω J t οσ ₍ as a limit, also a Tv ~ ^V W.Í«M ^V (ÉísLíOsí ) ~ ^V tíM«!s$$$v as a limit, and > ..«.rC: - fo.aiM U-Ui^sö) ™ fei.a:síA ⁿ ! as a barrier, then the starting node. and the rib arc search starting points (qm;) connected with graph edges, assigning costs (k) according to the edges between pairs of all further possible nodes, a search starts from the starting node satisfying the equation (I <p(q _w ) ^μ discrete rib arc search for endpoints (q _and g) on the graph, the points (q) belonging to the path with the least cost give the result rib arcs, since it is not stipulated that the last point along the path must be a lower arc, therefore, if the path ends with an upper arc , then the path is invalid, it is deleted, the final output of the procedure is the lower and upper rib arches (Jl0), which is connected to the input of the selector (V). where the bone contour refinement procedure based on dynamic programming (DPACSPE), whose input is connected to the lower and upper rib arches-1 (J12) appearing at one output of the selector (V), the series of pixels (p;) defining the starting curve (g) given by the rib points with a series of points DPACSPE input parameters (Π3) coming from the parameter setting unit (PBE). as the distance of the dividing points is 0), the value range of the displacement parameter (m), the cost of the step (K), the cost of moving to the side is 0.0, the value of the image intensity in py points (Kpi y)), the procedure directly saved the modified curve (g *) defining series of points (pg*), i.e. the costal arch, the steps of the solution: • by dividing the starting curve (g) into n parts, the i-th point (pi) of the starting curve, ♦ at every i-th point of the starting curve (g) setting perpendicular lines to the tangent of the curve at the given point, • on the perpendicular lines in two directions, shifting parameter value range (m) with number offset 0) acquisition of points, * the j-th shifted point of the pi point (pH), 47/43 ♦ image intensity value (I(py)) in the jth shifted point (p^) of the pt point, • copy image intensity values (Kpy)) into an intensity array CífPij)). • normalizing the elements of the intensity array (T(p _b jj) and then copying the result into the normalized intensity array (T'(pjj)), the elements of the latter array are the normalized intensity values (Ffpy)), • the values of the normalized intensity array (T '(pj)) modification and filtering based on a given rule system, * implementation of two types of filtering, normal and tangential filtering. ♦ when filtering in the normal direction, a new normalized intensity value (r*(pij)) filtered in the normal direction for any point (py) is formed by a weighted linear combination of the original normalized intensity values (Γ(ρ^)) in the vicinity of the point, • the output filtered values copying the filtered, normalized, intensity into the array (Τ' ^(pij)), ♦ the elements of the filtered, normalized, intensity array (T'*(pij) ) are then tangentially filtered, the results are in the twice filtered, normalized, intensity array (Τ'**(píj)) are stored, ♦ the order of the filters - normal direction and tangential direction - can be changed, ♦ a weighted directed graph is generated from the points (p) as follows kcfPLtj Ρψι) - Ttp^d + k^, Wpmj -* pm) ^::: T(pü)> MPmj “* Ρψ·0 “ HlvO * ' ^{with the} step cost (k _c ), the page step cost (k _cs ), and the array containing the pixel intensities (Tfpij)) as with image cost, * recording starting point (p.<), from here, starting the search k _c (pj —* po^í) ” According to HPeJ, the offset using the value range of the parameter (m) in both directions, • taking the end point (p _n ), where the number of points of the curve (n) ak _c (p _n nj p»/0 ™ ^: 0 based on the value range of the displacement parameter (m) in both directions, ♦ the output of the procedure is given by the path with the least cost leading from the starting point (pQ) to the end point (p _n ), the search is possible moving forward by position, because if all the accumulated costs (k<) are available up to the (i~l)~th position, then the accumulated cost can also be calculated in the i-th position based on the relationship (d(pij) a J<dPy) - min {k _s (pMA') +kv(pm· “* $(Ρυ)), jl<k<j+ l • the process can be parallelized, since the cost can be calculated independently for each shift (Ö) in a given position (p). 47/44 • determination of the optimal path: selection of the points (p) belonging to the last position with the lowest accumulated cost, then moving back along the neighbors with the lowest cost, the output of the procedure is the final rib ages (VBl), where the double bone contour based on dynamic programming refinement procedure (DPA2CSPE), the input of which is connected to the lower and upper rib arcs-2 (J11) appearing at the other output of the selector (V), the initial lower curve (g _a s _S ö) and initial upper curve (gíy _S ö) given by the rib points with a series of points ), the i-th point of the starting lower curve (p^ 4 ^the i-th point of the starting upper curves (p·^;), the DPA2CSPE input parameters from the parameter setting unit (PBE) (J14K as the distance of the division points (δ), the displacement parameter value range (m), step cost (k _c ya smoothness cost (k _es ; _fflasS g), image cost the procedure directly saved the series of points (ρ'&Μ arc) defining the modified upper curve (g*f _e aő) and the modified lower curve (g*^) megh replicating point series (p*^ íA i.e. ^the modified pair of rib arches, the steps of the solution: • dividing the initial lower curve (g^) and the initial upper curve (gty^) into n parts, the i-th point of the initial lower curve (p^ 0, the i-th point of the initial upper curves Φγ^ϊΧ • the initial curves setting lines perpendicular to the tangent of the curves at each i~edlk point, • recording points with an offset (ö) in the value range of the displacement parameter in two directions (m) on the perpendicular lines, • the i~jth shifted point of the i-th point of the starting upper curve φ^ arc) , the j~th shifted point of the i-th point of the starting lower curve (p^ ♦ in the j~th shifted point of the i-th point of the starting upper curve arc) the image intensity value (I(píu$íi q)F ^{of the} starting lower curve i In the j-th shifted point of the -th point (ρ ₃ ι»ά u) the image intensity value (Kpaus ü)k * copy the previous intensity values into the upper curve intensity array (T(pfe _is $ j)) and into the lower curve intensity array <Τφ«^ ij)), the upper curve intensity array (Tfp^ ij)). and normalizing the elements of the lower curve intensity array (Τφ»Ηό -j), then copying the results into the upper curve normalized intensity array ((T'(pfebö u))> and the lower curve into the normalized intensity array (T(p«is6 latter arrays elements are the normalized intensity values of the upper curve (P(pmvi arc)) and the normalized intensity values of the lower curve (Γ(ρ^ olk 47/45 • modification and filtering of the elements of the upper curve normalized intensity array ((T'fpíy^ ^)) and the lower curve normalized intensity array (Τ'φ^ íj)) based on a given set of rules, • implementation of two types of filtering, normal directional and tangential filtering, * when filtering in the direction of the norms, a new, normalized intensity value (F*(pe _S ö u)) corresponding to any starting upper curve point (pfctss y) in the direction of the norms of the upper curve (F*(pe S ö u)) is the normalized intensity of the original upper curve in the vicinity of ^the point values (F(pfei _$ $ ej)) are formed by a weighted linear combination, the resulting filtered values are stored in the filtered, normalized, intensity array (Τ^φΓ^βφ) of the upper curve, • when filtering in the normal direction for any starting lower curve point (p^ _{1 :S} ) are formed by a weighted linear combination of the normalized intensity value (T*(p ₈ kó φ)) of the original lower curve normalized intensity values (Fípus* ΐφ) located in the vicinity of the point, the output hair values are stored in the filtered, normalized, intensity array of the lower curve φ), • tangential filtering is then performed on the elements of the upper curve, normalized, intensity array (Τ^φ^ φ), the results are in the twice filtered, normalized, intensity array of the upper curve are stored (Τ'^^φ^ηό ij)), * on the elements of the filtered, normalized, intensity array of the lower curve (Τ' *φώ«ό φ) then a tangential filter is performed, the results of the lower curve are twice filtered, normalized, intensity are stored in an array (T'^pa^j)), ♦ the order of the filters (normal direction and tangential direction) can be changed, • the search for the optimal path takes place simultaneously on the starting lower curve (g^) and on the starting upper curve, thus for every position ( i), two displacement values, displacement of the upper curve (Sí), displacement of the lower curve (s ₂ ), belong to ♦ the displacement of the i-th point of the upper curve (sápi)) and the displacement of the i-th point of the lower curve (Βφφ must be fulfilled ai(s,(pj-s _; (p _; .. _s )i<l(65) and al(s ₂ (p _i )'-s _í (p _; „ _s )|< I conditions, * the steps are associated with costs, the step cost (k<) a consists of smoothness cost (k^as^J and the image cost (kc _W ), k _t (p _w -» P.) -> pJ + k^Cp;), ♦ the smoothness! cost (kcstmacity) 8 page access cost ( k _cs ) and from the step cost that moves the two curves apart (Εφ stands, i.e. Pi) Mfe-í Pi) - MPm “> Pi) > B Bi ss B 47/46 • the cost of moving to the side (1«) and the cost of the step that moves the two curves apart (ked) can be written as the average displacement (whey) (whey ~Sft- s>)/2>, the average step difference (davo) is by using the step weighting factor (y _s ) and the step weighting factor (γ^) that moves the two curves apart k _c Jp _w ρ _έ ) Y _s í s _AW -l)h ”>P>ü Μ _Αν0 (ρ _; )~ <1 _Λν0 (Χ ~ In Icais, ♦ the weighting factor for moving to the side (y _s ) and the weighting factor for moving the two curves apart (γ*) must be set in the interval 0...1, 0.2 can be used for both weighting factors, • the image cost is derived from the upper curve intensity array (T(p _fcSs$ y)) and the lower curve intensity array (Τφ^^)) according to the relationship ak^(p _; ) - f ΗΧρ^,ρΛίρ^,) f ♦ the ideal form of the function f is /(c _f , c,) ^::: 1 — (l ··· c,X1 - c / ), • after the normalization of the image cost, the path with the least cost is determined moving forward by position, • the accumulated cost <k _s (p0) for one cell of the search array is of the form min k^qj-ik.fp,), where 1(85(^)-8.^...)^1 and ^(p, ) - S2(pm)I<1 denotes /¾ points satisfying the relations, ♦ during the route search, the cost of the lowest-cost route belonging to the given point must be stored, • the optimal route can be obtained: among the points belonging to the last position, the smallest accumulated cost, and then moving back to the neighbors with the lowest cost, the procedure exits the final rib days (VBI). 2. The method according to claim 1, characterized by the fact that during the process of the convex circular arc-shaped edge-finding unit (ΚΚΛΕΚ), the value of the tilt angle («) of the matching circular arc is limited by using the average direction field, so the entire circular direction field is determined by incorrect determinations, basically the the effect of anatomical noises is reduced. 4//4/ 3. The method according to claim 1, characterized by the fact that during the process of the convex circular arc-shaped edge search unit (KKAÉK), the smoothed direction field formed from the lower rib arches is used when filtering the upper rib arches, while the smoothed direction field formed from the upper rib arches is used when filtering the lower rib arches. 4. The method according to claim I, characterized by the fact that the image prefilter unit (ESZÜ) contains a conical kernel, the structure of which is 0 20~r)/(r-n5) r 10 -rl.05r Ή OOOgauss...(r) ha |r|< 15, otherwise 0, the input of the kernel is the pre-segmented image (Jl ), the output is the pre-filtered image (H), the kernel receives the ESZÜ input parameters (32) from the parameter setting unit (PBE), which are the radius of the kernel ( r), the position (μ) and the standard deviation (0),